Introducing a Novel Figure of Merit for Evaluating Stability of Perovskite Solar Cells: Utilizing Long Short-Term Memory Neural Networks

This study introduces a novel figure of merit for evaluating the stability of perovskite solar cells (PSCs) by employing advanced Long Short-Term Memory (LSTM) neural networks to investigate degradation mechanisms. By harnessing the power of artificial intelligence and data analytics, we analyzed extensive datasets encompassing PSC parameters, experimental results, and environmental conditions, revealing critical insights into the degradation patterns affecting cell performance over time. Our findings indicate that the LSTM model effectively captures and predicts the complex relationships between key design parameters—efficiency, fill factor, and open-circuit voltage—and degradation-induced changes in PSCs. Specifically, we identified three degradation coefficients associated with the electron transport layer, hole transport layer, and perovskite active layer. These coefficients serve as a new figure of merit, facilitating numerical studies on degradation and stability in PSCs, mainly focusing on cesium lead halides. Furthermore, the enhanced LSTM architecture, featuring deeper layers, dropout for regularization, and batch normalization, demonstrated improved stability and training speed, leading to a test Mean Absolute Error (MAE) of 0.0354 and an R2 value of 0.9991, indicating near-perfect predictive accuracy. The comparative analysis of model complexity confirmed that increasing the sophistication of the LSTM model significantly enhances predictive accuracy and generalization capabilities. Identifying crucial design parameters offers actionable insights for optimizing PSC designs, materials selection, and operational conditions, ultimately contributing to enhanced long-term stability and efficiency of PSCs. Future research should prioritize using experimental datasets to achieve more realistic predictions, thereby driving innovation and unlocking the full potential of machine learning and deep learning in optimizing PSC design and performance.